Write JSON manifest file to disk

[beckjake]

There are a two significant components to this PR:

• Converted all the contracts to use json schemas instead of voluptuous and remove that dependency (through commit c9ae8d4)
• Converted the unparsed and parsed stages to use the APIObject subclasses instead of dictionaries and pass those around, and generate a manifest object (defined in contracts/graph/parsed.py) that can be converted to a json-serializable dict for writing to disk, and to the flat_graph required by the compiler.

Some side effects of note:

• I ended up removing the never-used all validator in validation_utils and writing a new any one.
• I had to restructure some of the parsing to pass extra parameters separately from the UnparsedNode object, so dbt.parser.parse_node had some slight signature changes, and dbt.parser.parse_seed now returns the unparsed node and the agate tables separately
• Artifact nodes are now allowed by the contract, otherwise I'd have had to do something awful with parse_node to make it play nice with UnparsedNodes
• I ended up changing the dbt.loader.GraphLoader a bit to only register node loaders, since macros have to be loaded before any of the nodes can be loaded anyway.

We probably want to understand/examine the performance of serializing all this JSON to strings and writing that out like the current code does. It might be good to make writing the manifest an option depending upon how slow it is.

This PR looks quite a bit bigger than it really is, a large chunk of it is converting sets/tuples to lists and stuff like that so the json schema validator is happy.

[cmcarthur]

this looks great! since we paired on most of it, and it's a big change, i'm going to give @drewbanin a chance to review before merge.

[drewbanin]

One question for my edification, but this looks A+

[drewbanin]

what's the deal with extra_ctes?

[beckjake]

They're stored as an OrderedDict, which doesn't have any nice JSON representation (the standard technique I've seen is to use a list of dicts like [{'key': 'foo', 'value': 'bar'}, {'key': 'baz', 'value': 'quux'}, ...]). So to store it as schema-validated JSON, we'd end up spending a lot of time packing/unpacking everything into/from the OrderedDict/list-of-dicts.

A better longer-term path is probably to convert the extra_ctes into just an array, or leave it out of the contract, like I did with the agate tables. I briefly tested out the array change and broke a lot of tests, so obviously there's something I don't understand going on.